Gemini 4 Carbon Tops Artificial Analysis Ranking With 59 Points, but the Model Is Closed
Google has shown Gemini 4 Carbon. On the independent Artificial Analysis Intelligence Index v5, a composite score built from tests on logic, code and science, it scored 59 points and took first place. On SEC-bench Pro, which tests finding vulnerabilities in others' code, it reached 88.2%. The context window extends to 1 million tokens, enough for several large repositories in a single request. Access is closed, so it is too early to build the model into product plans.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Google announced Gemini 4 Carbon, and the announcement is an odd one: the numbers are loud, but almost no one can actually try the model. If you are choosing a model for a product or an agentic system, it is important right now to separate three things: what Google has already claimed, what has been confirmed by an independent ranking, and what you will have to wait for until access opens.
What 59 points and first place mean
Artificial Analysis is an independent platform that compares models from different companies and rolls the results into one composite index. This is not Google's own report, so for the market this signal carries more weight than an ordinary press release. In version v5 of the index, Gemini 4 Carbon scored 59 points and placed first. Put simply, it is the new leader on overall "intelligence."
But a composite index does not tell you how good the model is for your specific task. It is like a grade point average: nice to have, but it does not tell you who will fix your particular bug better.
Code, vulnerabilities and long tasks
Here is what is known about its capabilities. The context window extends to 1 million tokens. That is roughly several large repositories, or an entire folder of contracts, that can be loaded into a single request without being split into pieces. Google also claims a significant gain in programming, logical reasoning and scientific research, and says the model can work for several days on a complex agentic task. Imagine assigning an AI to analyze a large project and not checking on it every half hour.
The most practical figure is 88.2% on SEC-bench Pro. This is a benchmark for finding vulnerabilities in software: the model is given code and must find the dangerous spots in it. The higher the percentage, the more holes it spots. For teams doing code audits, this matters more than the overall index.
Yet Google does not hold first place in coding. In the ranking of AI programming agents, Carbon takes second place behind Claude Opus 5.5 from Anthropic. If you use AI as a coding partner, the leader has not changed: reasoning well in general and carefully editing code in a large project are different skills.
Pricing and access
API rates:
- input: $2 per 1 million tokens;
- output: $10 per 1 million tokens;
- cached input: $0.10 per 1 million tokens.
An important caveat: these are launch prices with a 50% discount, and they will later rise to $4 for input and $20 for output. Let us work through an example. A request with 100 thousand input tokens and a 5 thousand token response will cost about $0.25 now and about $0.50 after the increase. If a large document repeats between requests and is served from the cache, 100 thousand input tokens cost about $0.01. For agents that reread the same context again and again, caching becomes the main source of savings.
Access is still closed. Google is opening Carbon to selected cybersecurity specialists through the Fairwind Program. The model is promised for the Gemini app and AI Studio, but there are no precise dates.
What to do in practice
First: do not build Carbon into product plans until a public API appears. For now everything rests on Google's claims and the Artificial Analysis ranking; independent checks on your own tasks will only be possible once access opens.
Second: if you have a lot of repeated context, budget against the future $4/$20 rate, not the launch rate. The discount is temporary, and costs calculated on it will double over time.
Third: if you work in security, watch the Fairwind Program. That is where the model will appear first, and 88.2% on SEC-bench Pro is a reason to take a closer look.
The limitation of the announcement is obvious: Google reported results but did not let anyone verify them. While the model is unavailable to the broad audience, the ranking and benchmarks remain a promise. Carbon's real value will become clear once developers run it on their own repositories and compare it with Claude Opus 5.5 in working scenarios.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What 59 points and first place mean?
Artificial Analysis is an independent platform that compares models from different companies and rolls the results into one composite index. This is not Google's own report, so for the market this signal carries more weight than an ordinary press release. In version v5 of the index, Gemini 4 Carbon scored 59 points and placed first. Put simply, it is the new leader on overall "intelligence." But a composite index does not tell you how good the model is for your specific task.
What are the key details on Code, vulnerabilities and long tasks?
Here is what is known about its capabilities. The context window extends to 1 million tokens. That is roughly several large repositories, or an entire folder of contracts, that can be loaded into a single request without being split into pieces. Google also claims a significant gain in programming, logical reasoning and scientific research, and says the model can work for several days on a complex agentic task. Imagine assigning an AI to analyze a large project and not checking on it every half hour.
What are the key details on Pricing and access?
API rates: input: $2 per 1 million tokens; output: $10 per 1 million tokens; cached input: $0.10 per 1 million tokens. An important caveat: these are launch prices with a 50% discount, and they will later rise to $4 for input and $20 for output. Let us work through an example. A request with 100 thousand input tokens and a 5 thousand token response will cost about $0.25 now and about $0.50 after the increase.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.